"GradientDef is not supported in reading Dataset related functions: ",
        text_proto);
    return {};
  }
  std::vector<UniqueFuncPtr> ret;
  for (const FunctionDef& fdef : fdef_lib.function()) {
    // Make a copy so that we can mutate it.
    FunctionDef fdef_to_load = fdef;
    if (mutate_proto_func) {
      (*mutate_proto_func)(&fdef_to_load);
    }
    VLOG(1) << "Adding func to graph: " << fdef_to_load.DebugString();

    int64_t target_id = target_tensor_ids[i];
    if (sources_set.find(target_id) != sources_set.end()) {
      auto tensor = targets[i];
      sources_that_are_targets.insert(
          std::make_pair(target_id, TapeTensor(tensor)));
    }
  }

  TF_RETURN_IF_ERROR(GradientTape::ComputeGradient(
      vspace, target_tensor_ids, source_tensor_ids, sources_that_are_targets,

  // in order to update them as appropriate.
  auto new_entry = std::make_shared<Block>();
  lru_list_.push_front(key);
  lra_list_.push_front(key);
  new_entry->lru_iterator = lru_list_.begin();
  new_entry->lra_iterator = lra_list_.begin();
  new_entry->timestamp = timer_seconds_();
  block_map_.emplace(std::make_pair(key, new_entry));
  return new_entry;
}

  auto file = std::make_unique<TF_RandomAccessFile>();
  std::string translated_name = TranslateName(fname);
  ops_->new_random_access_file(filesystem_.get(), translated_name.c_str(),
                               file.get(), plugin_status.get());

  if (TF_GetCode(plugin_status.get()) == TF_OK)
    *result = std::make_unique<ModularRandomAccessFile>(

    } else {
      bytes_transferred = 0;
    }
    TF_SetStatus(status, TF_OK, "");
    return bytes_transferred;
  };
  for (size_t block_size = 2; block_size <= 4; block_size++) {
    // Make a cache of N-byte block size (1 block) and verify that reads of
    // varying offsets and lengths return correct data.
    tf_gcs_filesystem::RamFileBlockCache cache(block_size, block_size, 0,

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(

  TFE_OpSetAttrString(op.get(), "container", "", 0);
  // Use the special GUID for no buffer sharing
  //
  // TODO(allenl): Should we provide a better API for this? AFAIK this is the
  // only reasonable way to make variables with no aliasing using the eager C
  // API.
  std::string no_sharing = "cd2c89b7-88b7-44c8-ad83-06c2a9158347";
  TFE_OpSetAttrString(op.get(), "shared_name", no_sharing.c_str(),

}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,
    const TF_Output* inputs, std::vector<OutputTensor>* input_tensors,

      TFE_TensorHandleDevicePointer(tensor, status));
  if (TF_GetCode(status) != TF_OK) return nullptr;
  if (parallel_tensor->num_tensors() == 1) {
    // Copy-off for single-device tensors is allowed to make debugging dynamic
    // control flow easier.
    return TFE_TensorHandleCopySharingTensor(parallel_tensor->tensor(0),
                                             status);
  } else {
    TF_SetStatus(